tests: fix NameError in daemon_sock / fake_daemon fixtures (macOS CI)

[danielhertz1999-bit]

The build & test (macOS) gate is currently red on main (incl. the Release v1.2.0 run). This fixes the most clear-cut part of it.

What's broken

daemon_sock (test_mcp_tools.py) and fake_daemon (test_socket_disconnect_reconnect.py) both reference tmp_path in the fixture body without it being a parameter:

@pytest.fixture(scope="module")
def daemon_sock() -> "Path":
    sock_dir = tmp_path / "sock"   # NameError: name 'tmp_path' is not defined

So every test using them errors at setup — 5 errors on the macOS gate. (daemon_sock is also scope="module", where the function-scoped tmp_path wouldn't be injectable anyway.)

Fix

Bind the control socket under a short tempfile.mkdtemp(prefix="iai-sock-") dir (the same pattern as the existing short_socket conftest fixture). This:

• removes the undefined-name reference, and
• keeps the AF_UNIX path under the ~104-char sun_path limit — pytest's macOS tmp_path (/private/var/folders/.../pytest-of-runner/...) overflows it, which is the same root cause behind the other macOS socket failures.

Both fixtures already rmtree their sock_dir on teardown, so cleanup is unchanged. Verified the modules still collect cleanly.

Scope / remaining

I kept this to the two definitely-broken fixtures (they NameError 100% of the time). The other macOS failures — test_concurrency, test_capture_transcript_no_spawn_defer, test_bridge_socket_first — share the same AF_UNIX-path-too-long root cause (they build tmp_path / "sock"), and pass on Linux today. Happy to route those sockets through short dirs too in a follow-up, but I didn't want to churn currently-green fixtures I can't exercise on a macOS runner locally — your call on whether you'd like that.

🤖 Generated with Claude Code